SIMD

Question 1

what is simd

Answer 1

more unit instrutions next to each other to do same op to different data

transistor budget
less power
and twice performance ideally

Question 2

what are the vector sizes in simd

Answer 2

SSE 128
AVX 256
AVX 512

Question 3

intrinsic functions

Answer 3

data types: __m256 (8x32), __m256d (4x64), __m256i (any)

vector length: _mm512, _mm256
function: add, mul, sub, load, store
type and precision: pd (packed double), ps (packed single), ss (scalar single)

loadu: unaligned
load: aligned

Question 4

vec efficiency

Answer 4

VE = N/vl

N trip count (how many times youre doing the vectorized loop)
vl vector lenght

Question 5

simd intrinsics cons

Answer 5

low level
error prone

thats why we use compiler generated simd code!!

Question 6

when does autovectorisation happen and when does it fail

Answer 6

• if possible
• if beneficial

fails for:
- data dependency
- alignment
- mixed data types
- function calls in loop

Question 7

can we parallelise or vectorzie a loop with loop carried dependency

Answer 7

parallelise no
vectorize yes: if the distance is greater than the vector size

Question 8

openmp simd

Answer 8

pragma omp simd

maybe a for loop right after if #pragma omp for simd

• can have private, firstprivate and rediction
• safelen(l) safe length to vectorise (max vec length)
• linear(list[:linear step])
  aligned(list[:alignment])

for loops: simd chunks
schedule(simd:static, 5) chooses roughly 5 chunks but favours alignment, no remained loops

-remainder: end
-peel loop: beginning